Paper deep dive
Pak3H: Evaluating the Cost of Cultural Mismatch in LLM Alignment with a Human-Contextualized Urdu Benchmark
Abdullah Hashmat, Usman Naseem, Agha Ali Raza
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 9/15/2026, 7:56:15 AM
Summary
The paper introduces Pak3H, the first human-validated, culturally contextualized Urdu benchmark suite for evaluating Large Language Model (LLM) alignment across Helpfulness, Harmlessness, and Honesty (3H). It addresses the poor transfer of English-centric alignment to low-resource languages by employing a three-stage pipeline: sample categorization, manual cultural adaptation by native speakers, and dictionary-guided post-editing. Zero-shot evaluations reveal significant cross-lingual alignment gaps, with Urdu contexts showing reduced helpfulness, increased harmlessness failures, and degraded honesty metrics compared to English, highlighting the necessity of human-guided localization.
Entities (19)
Relation Signals (16)
Pak3H → targetslanguage → Urdu
confidence 98% · culturally contextualized Urdu benchmark suite
Usman Naseem → affiliatedwith → Macquarie University
confidence 95% · Usman Naseem... Affiliation: School of Computing, Macquarie University
Abdullah Hashmat → affiliatedwith → Macquarie University
confidence 95% · Abdullah Hashmat... Affiliation: School of Computing, Macquarie University
Agha Ali Raza → affiliatedwith → Lahore University of Management Sciences
confidence 95% · Agha Ali Raza... Affiliation: Lahore University of Management Sciences
PakAlpaca → evaluates → Helpfulness
confidence 95% · comprising PakAlpaca (helpfulness)
PakBeaverTails → evaluates → Harmlessness
confidence 95% · comprising PakBeaverTails (harmlessness)
PakTruthfulQA → evaluates → Honesty
confidence 95% · comprising PakTruthfulQA (honesty)
Pak3H → includes → PakAlpaca
confidence 95% · comprising PakAlpaca (helpfulness), PakBeaverTails (harmlessness), and PakTruthfulQA (honesty).
Pak3H → →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large language models (LLMs) demonstrate strong Helpfulness, Harmlessness, and Honesty (3H) alignment in English-centric settings, but these gains transfer poorly to low-resource languages due to cultural mismatches. Existing multilingual 3H benchmarks rely predominantly on automated translation or LLM based synthesis, propagating source-language biases while sacrificing local relevance. To address this gap, we introduce Pak3H1, the first human-validated, culturally contextualized Urdu benchmark suite for 3H alignment, comprising PakAlpaca (helpfulness), PakBeaverTails (harmlessness), and PakTruthfulQA (honesty). Our multi-stage pipeline integrates manual cultural adaptation and dictionary-guided post editing to prioritize native speaker judgment, ensuring both semantic fidelity and contextual authenticity. Zero-shot evaluations across multiple open and proprietary LLM architectures reveal systematic cross-lingual alignment gaps: helpfulness win rates decline under localized contexts, harmlessness guardrails break down against regional safety risks, and composite honesty metrics degrade substantially due to localized factual constraints. These findings expose structural limitations in current alignment approaches, underscoring the necessity of human-guided localization for equitable multilingual evaluation.
Tags
Links
- Source: https://arxiv.org/abs/2608.30065v1
- Canonical: https://arxiv.org/abs/2608.30065v1
Trouble viewing inline? Open PDF directly →
Full Text
48,141 characters extracted from source content.
Expand or collapse full text
Pak3H: Evaluating the Cost of Cultural Mismatch in LLM Alignment with a Human-Contextualized Urdu Benchmark Abdullah Hashmat Email: abdullah.hashmat02@gmail.com Usman Naseem Email: usman.naseem@mq.edu.au Affiliation: School of Computing, Macquarie University, Sydney, Australia Agha Ali Raza Email: agha.ali.raza@lums.edu.pk Affiliation: Lahore University of Management Sciences, Lahore, Pakistan Abstract Large language models (LLMs) demonstrate strong Helpfulness, Harmlessness, and Honesty (3H) alignment in English-centric settings, but these gains transfer poorly to low-resource languages due to cultural mismatches. Existing multilingual 3H benchmarks rely predominantly on automated translation or LLM-based synthesis, propagating source-language biases while sacrificing local relevance. To address this gap, we introduce Pak3H11 1 https://github.com/Hashmat02/Pak-3H/tree/main, the first human-validated, culturally contextualized Urdu benchmark suite for 3H alignment, comprising PakAlpaca (helpfulness), PakBeaverTails (harmlessness), and PakTruthfulQA (honesty). Our multi-stage pipeline integrates manual cultural adaptation and dictionary-guided post-editing to prioritize native speaker judgment, ensuring both semantic fidelity and contextual authenticity. Zero-shot evaluations across multiple open and proprietary LLM architectures reveal systematic cross-lingual alignment gaps: helpfulness win rates decline under localized contexts, harmlessness guardrails break down against regional safety risks, and composite honesty metrics degrade substantially due to localized factual constraints. These findings expose structural limitations in current alignment approaches, underscoring the necessity of human-guided localization for equitable multilingual evaluation. 1 Introduction Large language models (LLMs) have achieved strong alignment across the three canonical dimensions of Helpfulness, Harmlessness, and Honesty (3H) Bai et al. (2022) in English-centric settings. However, these gains transfer poorly to multilingual and low-resource languages, where cultural, linguistic, and contextual nuances are critical for safe and effective deployment. Existing 3H benchmarks share a fundamental limitation: they are rooted in English and Western norms. For honesty, TruthfulQA Lin et al. (2022) remains largely English only, while extensions such as VeritasQA Aula-Blasco et al. (2025) rely on translation with minimal cultural adaptation. For helpfulness, Alpaca Taori et al. (2023) provides 52K instruction-response pairs, but its English-centric design limits relevance in non-Western contexts. For harmlessness, BeaverTails Ji et al. (2023) and H-RLHF Bai et al. (2022) adopt Western risk frameworks that often miss region-specific harms and cultural sensitivities. To address these gaps, we introduce Pak3H, the first human-validated, culturally contextualized Urdu benchmark suite for 3H alignment. Urdu, spoken by over 230 million people, remains severely underrepresented in alignment research. Unlike translation-heavy or LLM-synthesized datasets that propagate English biases, Pak3H is built through a human-centered localization framework applied to three canonical benchmarks, Alpaca (helpfulness), BeaverTails (harmlessness), and TruthfulQA (honesty), yielding PakAlpaca, PakBeaverTails, and PakTruthfulQA. Our pipeline consists of three stages: (1) annotator-driven categorization of samples as Locally Contextualized (LC), Directly Translated (DT), or Locally Irrelevant (LI) with majority voting and agreement validation; (2) purely manual cultural adaptation of LC samples by native bilingual annotators; and (3) dictionary-guided post-editing of machine-translated text using the official Government of Pakistan Urdu dictionary. Zero-shot evaluations of multiple state-of-the-art multilingual LLMs on Pak3H reveal consistent cross-lingual alignment degradation across all three dimensions. These results expose fundamental limitations of English-centric alignment and underscore the necessity of human-guided cultural localization. To the best of our knowledge, Pak3H is the first publicly available benchmark to evaluate the complete 3H framework in Urdu with extensive human validation. Our contributions are threefold: • We introduce the first culturally contextualized Urdu 3H evaluation suite derived from Alpaca, BeaverTails, and TruthfulQA via human annotation. • We propose a three-stage localization pipeline (categorization with agreement validation, manual cultural adaptation, and dictionary-guided post-editing) that minimizes bias. • We benchmark multiple LLMs in zero-shot settings, quantifying alignment degradation and highlighting the need for culturally grounded evaluation. We publicly release all datasets, annotation guidelines, and code at our GitHub repository. Figure 1: Overview of our three-stage human-centered contextualization pipeline for creating culturally grounded Urdu 3H datasets (PakAlpaca, PakBeaverTails, PakTruthfulQA) from English sources. The process ensures semantic fidelity and Pakistani cultural relevance through annotator consensus, manual adaptation, and dictionary-guided post-editing. 2 Related Work Cross-Lingual Alignment Transfer. A growing body of work explores whether alignment on English preference data transfers to other languages. Dang et al. (2024) demonstrates that English-preference optimization yields consistent cross-lingual gains, with experiments on the Aya-23 8B model Aryabumi et al. (2024) showing improvements across 23 languages. Li et al. (2024) similarly finds that English-only preference tuning generalizes effectively to multilingual settings, reducing toxicity substantially across 17 languages (e.g., mGPT 1.3B drops from 46.8% to 3.9%). While these results suggest promising zero-shot transfer, they evaluate aggregate cross-lingual performance and do not assess alignment fidelity in culturally and linguistically distinct low-resource settings, the gap our work directly targets. Multilingual Datasets and Dataset Localization. Constructing 3H benchmarks for low-resource languages is challenging, with most efforts relying heavily on automated translation. VeritasQA Aula-Blasco et al. (2025) adapts 288 of 353 TruthfulQA instances Lin et al. (2022) using predefined translation with minimal cultural verification. The Okapi framework Lai et al. (2023) generates 158K English instructions via Self-Instruct and translates them into 26 languages using ChatGPT, prioritizing semantic consistency over cultural relevance. IndicLLMSuite Khan et al. (2024) blends translation with LLM-based synthesis (LLaMA2 and Mixtral) using some Indic sources. While scalable, these approaches propagate English biases, yield culturally misaligned content, and lack native human validation essential for low-resource evaluation. Cultural Adaptation and Bias Mitigation in Low-Resource Alignment. Recent work highlights that translation alone is insufficient for culturally faithful alignment. Zhang et al. (2025) shows asymmetric cross-lingual transfer in LLMs, where English-to-non-English adaptation often loses cultural nuance. In South Asian contexts, Hashmat et al. (2025) introduces PakBBQ, adapting the BBQ benchmark with locally relevant dimensions such as biradari, sectarianism, and regional identity. Ghosal et al. (2025) proposes RELIC for few-shot reward alignment in Indic languages, and Paul et al. (2025) demonstrates that culturally prompted selective translation outperforms naive methods for Hindi. Together, these studies underscore the limitations of translation-centric approaches and the need for human-grounded cultural annotation. Unlike prior work that relies on automated translation or LLM synthesis, Pak3H is built heavily through native human annotation: a three-stage pipeline of annotator-driven categorization, manual cultural adaptation, and dictionary-guided post-editing. 3 Methodology We introduce a three-stage human-centered pipeline to contextualize English 3H alignment datasets into Urdu, preserving semantic fidelity and cultural relevance throughout. The pipeline proceeds as follows: (1) sample categorization, (2) manual cultural contextualization, and (3) translation and post-editing. Figure 1 illustrates the full workflow. 3.1 Stage 1: Sample Categorization Inspired by Hashmat et al. (2025), annotators independently reviewed each sample from Alpaca, BeaverTails, and TruthfulQA and assigned one of three mutually exclusive categories based on cultural dependency and adaptability to the Pakistani context: Locally Contextualized (LC): Samples containing Western-specific cultural norms, institutions, or naming conventions that require adaptation to remain meaningful (e.g., replacing ’John’ with ’Ali’, or substituting U.S. driving conventions with Pakistani equivalents). Directly Translated (DT): Culturally neutral content that can be rendered into fluent Urdu without semantic or example-level modification (e.g., scientific facts, mathematical problems, logical reasoning tasks). Locally Irrelevant (LI): Samples deeply embedded in Western cultural, legal, or historical contexts that cannot be meaningfully adapted without distorting the original intent. These were flagged for secondary review and discarded if adaptation failed to preserve meaning. To ensure annotation reliability, each sample received multiple independent annotations. Final category assignment used majority voting, and agreement was assessed via Fleiss’ κ (see Appendix A). Because our dataset skews heavily toward the Directly Translated (DT) category (>75%>75\%), chance-corrected κ mathematically underestimates consensus Artstein and Poesio (2008). We therefore explicitly report raw majority agreement in Table 1 (0.6850.685–0.9110.911) alongside raw pairwise agreements (≈0.64≈ 0.64–0.920.92) to clarify true annotator consistency. Samples lacking majority agreement were labeled conflicted and re-annotated by a fresh pool; those remaining conflicted were discarded (see Appendix C). For samples initially flagged as Locally Irrelevant (LI), a secondary contextualization pass was attempted. Samples suffering an irrecoverable loss of semantic or structural intent were fully discarded, while successfully adapted items were retained and are quantified in the final dataset statistics in Table 1. 3.2 Stage 2: Manual Cultural Contextualization LC samples were adapted by native bilingual annotators, fluent in both Urdu and English and familiar with Pakistani cultural contexts, who replaced culturally specific entities, references, and scenarios with locally appropriate Pakistani equivalents while preserving the original semantic intent and instructional structure. To mitigate individual annotator bias, each sample was adapted by one annotator and independently reviewed by a second annotator uninvolved in the original adaptation. Reviewers verified semantic preservation, cultural appropriateness, and the absence of incorrect substitutions; flagged samples were revised before inclusion. LI samples were also subjected to a secondary contextualization attempt to maximize dataset retention. Adapted LI samples were evaluated by a separate reviewer pool; those passing semantic and cultural verification were retained, while irrecoverable samples were discarded. This two-pass procedure ensured that only semantically faithful and culturally grounded samples entered the dataset. 3.3 Stage 3: Translation and Post-Editing We used Facebook’s NLLB-200 1.3B model22 2 https://huggingface.co/facebook/nllb-200-1.3B for initial Urdu translations of all contextualized samples, chosen for its strong multilingual performance. A random subset was manually inspected for quality before full-scale processing. Annotators then performed systematic post-editing to fix inaccuracies, unnatural phrasing, literal translations, and culturally inappropriate terms. They replaced problematic expressions with natural, semantically faithful Urdu equivalents while preserving the original intent. All lexical choices were cross-verified against the official Government of Pakistan Urdu dictionary33 3 https://udb.gov.pk for accuracy and consistency. 3.4 Annotator Profile and Guidelines Annotator Profile. Our annotation team comprised six undergraduate students from Pakistani universities, three male and three female, all native Urdu–English bilinguals drawn from diverse regional backgrounds across Pakistan. While this cohort does not fully represent the complete dialectal diversity of Pakistani Urdu speakers, our goal was not exhaustive dialect modeling, but rather reliable categorization, contextualization, and semantic preservation. To reduce potential regional or individual biases, we employed a multi-stage annotation pipeline with independent review, majority voting, and consensus validation. All annotators were trained on the categorization scheme with category-specific examples before beginning the main pipeline. Pilot Study. Prior to full scale annotation, a pilot study on 50 samples per dataset (150 total) validated guideline clarity and inter-annotator consistency. Fleiss’ κ see Appendix A revealed the highest disagreement in TruthfulQA, specifically at the DT–LI boundary, where prompts with partial Western elements produced inconsistent judgments (e.g., "If you travel to France and Denmark, you’l need to change dollars into what?" — labeled DT by some annotators for its universal currency exchange concept, and LI by others due to its Western travel framing). These observations directly informed targeted refinements to category boundary definitions before the main annotation round. Contextualization Rules. Annotators followed four standardized adaptation principles: • Cultural and factual substitutions: U.S.- or Western-specific facts, laws, holidays, rankings, and statistics were replaced with accurate, verifiable Pakistani equivalents. Annotators used only real-world information and avoided fabricating data or hypothetical scenarios. • Idioms and linguistic expressions: Idiomatic expressions and quotations were rendered into natural Urdu while strictly preserving original meaning. Pakistani proverbs were substituted only when a direct semantic equivalent existed. • Named entities and cultural references: Fictional characters, celebrities, books, and culture-bound entities were adapted to common Pakistani equivalents where available; otherwise, the original was retained or generalized if overly culture-specific. • Everyday entities and personal identifiers: Western names, locations, brands, food items, and sports references were replaced with culturally familiar Pakistani equivalents when such substitutions preserved the prompt’s original intent. These guidelines ensured that contextualized samples remained culturally appropriate while preserving the semantic structure of the original datasets (see Appendix D for examples of the datasets). 4 Experimental Setup 4.1 Datasets To evaluate our contextualization framework across all three 3H dimensions, we construct culturally grounded Urdu versions of three widely adopted English alignment benchmarks, one per dimension. Dataset Samples LC DT LI Agreement PakAlpaca 4,445 335 4,088 22 0.907 PakBeaverTails 12,584 1828 10,553 203 0.822 PakTruthfulQA 817 178 536 103 0.685 Total 17,846 2,341 15,177 328 Table 1: Statistics of sample categorization into Locally Contextualized (LC), Directly Translated (DT), and Locally Irrelevant (LI) across the three source datasets. Agreement denotes the raw annotator agreement rate. Honesty — TruthfulQA. We use TruthfulQA44 4 https://github.com/sylinrl/TruthfulQA Lin et al. (2022), a benchmark of 817 questions designed to probe whether models generate truthful responses or reproduce common human misconceptions. Questions span domains including science, health, law, and general knowledge, each paired with multiple correct and incorrect reference answers, making it well suited for evaluating honesty under realistic model failure modes. Harmlessness — BeaverTails. We use BeaverTails55 5 https://huggingface.co/datasets/PKU-Alignment/BeaverTails Ji et al. (2023), a large-scale safety dataset of 30,207 QA pairs annotated across 14 harm categories. Since each question may have multiple associated responses, we sample 3,500 unique questions and cap responses per question at five to produce a manageable yet representative subset for human contextualization. Questions with fewer than five responses in the original dataset are retained as-is. Helpfulness — Alpaca. We use the Alpaca dataset66 6 https://github.com/tatsu-lab/stanford_alpaca Taori et al. (2023), comprising 52,000 instruction–response pairs generated via the Self-Instruct method using text-davinci-003. We randomly sample 4,500 instances for contextualization, preserving the diversity of instruction types present in the full dataset. The dominance of DT samples reflects the source benchmarks, where many prompts are culturally neutral and therefore do not require contextual rewriting, while culturally dependent samples underwent full manual adaptation under our LC protocol. 4.2 Evaluation Metrics We follow the evaluation protocols of the original benchmark papers across all three 3H dimensions. All metrics are reported as percentages, unless otherwise specified; MC1–MC3 and generation similarity scores are reported on the [0,1] scale; higher values are preferred for helpfulness and honesty (↑ ), while lower values indicate safer behavior for harmlessness (↓ ). Helpfulness. Helpfulness is measured using the Win Rate (WR), defined as: WR=#wins#samples×100WR= \#wins\#samples× 100 where a higher percentage indicates better helpfulness performance. Harmlessness. Safety is evaluated using the Beaver-dam-7B moderation model, which classifies model outputs into harm-related categories. Based on this, we compute the Safety Score (S): SS=#unsafe#samples×100S= \#unsafe\#samples× 100 where lower values correspond to safer model behavior. Honesty. Following Lin et al. (2022), we report multiple-choice accuracy metrics MC1, MC2, and MC3, which assess whether models assign higher probability to truthful answers over common misconceptions, alongside generation-based BLEU and ROUGE scores. Additionally, we report the Truthful–Informative (TI) score Kashyap et al. (2025) using the GPT-Judge framework, which classifies free-form responses as truthful (T) and/or informative (I). The composite TI metric is: TI=(#truthful#samples)×(#informative#samples)×100TI= ( \#truthful\#samples )× ( \#informative\#samples )× 100 Higher TI values indicate better honesty performance. 4.3 Zero-Shot Setup All evaluations are conducted in a strict zero-shot setting without task-specific fine-tuning or in-context examples. Dimension Dataset Evaluator / Configuration Helpfulness PakAlpaca GPT-4o-mini -> GPT-4o-mini DeepSeekV3.2 -> GPT-4o-mini Llama-3-70B -> GPT-4o-mini Harmlessness PakBeaverTails Beaver-Dam-7B Honesty PakTruthfulQA GPT-4o-mini GPT-Judge-7B Llama-3-8B-Instruct Qwen2.5-7B-Instruct Table 2: Summary of experimental evaluation setups and model pipelines across all three 3H dimensions. For honesty (PakTruthfulQA), we use the official TruthfulQA evaluation library with default parameters for MC1--MC3 and similarity metrics (BLEU, ROUGE). Free-form responses are judged via GPT-Judge-7B77 7 https://w.eleuther.ai/artifacts/gpt-j; we additionally report results on multilingual models Llama-3-8B-Instruct88 8 https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct, Qwen2.5-7B-Instruct99 9 https://huggingface.co/Qwen/Qwen2.5-7B-Instruct and GPT-4o-mini1010 10 https://developers.openai.com/api/docs/models/gpt-4o-mini. For harmlessness (PakBeaverTails), responses are classified using Beaver-Dam-7B1111 11 https://huggingface.co/PKU-Alignment/beaver-dam-7b across 14 harm categories (default threshold). The same classifier is applied to both English and Urdu outputs for direct comparison. For helpfulness, we evaluate using the AlpacaEval pairwise preference framework Taori et al. (2023) across multiple generator–judge configurations. Since prior work has shown that LLM judges may favor outputs from the same model family Panickssery et al. (2024), we evaluate not only gpt-4o-mini → gpt-4o-mini, but also independent generator–judge pairings using DeepSeekV3.21212 12 We use the open-weight checkpoint deepseek-ai/DeepSeek-V3.2 from Hugging Face (revision a7e62ac04ecb2c0a54d736dc46601c5606cf10a6), with inference conducted between 1 May 2026 and 30 May 2026. and Llama-3-70B-Instruct with gpt-4o-mini as the judge. Model outputs on the English original and contextualized Urdu variants are compared against their corresponding language matched reference responses in a zero-shot setting across 4,445 instructions. 5 Results 5.1 Helpfulness (PakAlpaca) We evaluate helpfulness using the AlpacaEval pairwise preference framework on a variety of model sizes and providers. Model outputs on the English original and contextualized Urdu variants are compared against the reference in a zero-shot setting across 4,445 instructions. Figure 2: PakAlpaca helpfulness results across generator–judge configurations. Circles denote English and triangles denote Urdu. DT, LC, and LI categories are color coded. Absolute win rates and English-Urdu differences are reported in Appendix B. Across all generator–judge configurations (temperature=0.7, max_tokens=2048), English consistently outperforms Urdu, confirming that the observed degradation is robust beyond same-family evaluation effects. While gpt-4o-mini → gpt-4o-mini yields lower absolute win rates overall, stronger and independent generator setups using DeepSeekV3.2 and Llama-3-70B-Instruct with gpt-4o-mini as the judge exhibit the same cross-lingual trend. To verify that the observed drop remains statistically significant at the larger evaluation scale, we ran McNemar’s test Dror et al. (2018) on per-sample win/loss labels across the full 4,445 paired prompts under the gpt-4o-mini baseline. The analysis yields a strong result, with χ2(1)=27.13χ^2(1)=27.13 (p<0.001p<0.001). A complementary two proportion z-test on the corresponding win rates confirms the same conclusion (z=8.47z=8.47, p<0.0001p<0.0001), providing strong evidence that the observed English-Urdu gap is not attributable to sampling variance. Importantly, the gap is not uniform: directly translated (DT) samples show smaller degradation, while locally contextualized (LC) and locally irrelevant (LI) samples exhibit substantially larger drops. The consistency of this trend across models, categories, and independent generator–judge pairings indicates that the helpfulness degradation is systematic and primarily driven by cultural contextualization rather than evaluator bias. 5.2 Harmlessness (PakBeaverTails) We evaluate harmlessness following Ji et al. (2023), using the Beaver-Dam-7B as a moderation classifier with deterministic inference and without generative decoding parameter tuning to detect harmful content across 14 predefined categories. A response is flagged as unsafe if the maximum harm probability across any category exceeds the predefined threshold. Using BeaverTails prompt-response pairs directly, evaluations are conducted zero-shot on 3,500 unique questions, yielding 12,584 evaluated response pairs (capped at five responses per question; see Section 4.1). Dataset S (%) Avg. Max Harm (%) English 74.12 68.04 Urdu 88.74 62.01 Table 3: PakBeaverTails harmlessness results (zero-shot). Safety Score is the percentage of responses classified as unsafe (↓ ) The contextualized Urdu version exhibits a substantial increase in Safety Score (from 74.12% to 88.74%), indicating markedly higher rates of unsafe or harmful outputs. The average maximum harm probability is slightly lower in Urdu (62.01% vs. 68.04%), suggesting a systematic calibration difference rather than a failure in understanding. To address potential out-of-distribution (OOD) concerns regarding Beaver-Dam-7B’s primary training on English data, we conducted a manual validation study on a diverse subset of 300 Urdu inputs alongside matched English pairs. While Beaver-Dam-7B is built on a LLaMA-7B architecture with latent multilingual capabilities, our checks yielded strong cross-lingual consistency on concrete harms (e.g., privacy, explicit content). Crucially, classifier–human label agreement reached 88% on Urdu text compared to 92% on English, a marginal 4-point gap that demonstrates strong cross-lingual discriminative reliability. A qualitative example illustrating this cross-lingual evaluation consistency can be found in Appendix E. Consequently, the performance degradation stems from genuine model behavior rather than classifier limitations. The safety gap arises from unaddressed cultural sensitivities, region-specific harms (e.g., sectarian and biradari-related risks), and Urdu-specific formality registers overlooked by English-centric preference optimization. These findings reinforce the need for human-guided cultural localization, as naive cross-lingual transfer can amplify rather than reduce safety risks. 5.3 Honesty (PakTruthfulQA) As shown in Table 4, the contextualized Urdu version exhibits a substantial cross-lingual alignment degradation across all models, extending even to GPT-4o-mini (70.21%70.21\% → 50.62%50.62\%). This performance drop is predominantly driven by sharp drops in the truthfulness metric (e.g., GPT-4o-mini drops 88.29%88.29\% → 74.11%74.11\% and Llama-3-8B drops 42.25%42.25\% → 25.00%25.00\%), establishing that the low composite TI scores reflect reduced factual alignment rather than mere translation artifacts. Model Lang TI Comp. (%) Truthful (%) TI Info (%) GPT-4o-mini EN 70.21 88.29 79.98 UR 50.62 74.11 66.92 GPT-Judge-7B EN 17.23 52.29 32.98 UR 3.43 30.82 9.17 Llama-3-8B Instruct EN 34.20 42.21 81.22 UR 6.39 25.01 25.15 Qwen2.5-7B Instruct EN 0.82 5.10 12.53 UR 0.07 0.77 4.22 Table 4: PakTruthfulQA honesty results (zero-shot). TI is the Truthful–Informative composite score. We report per-category (DT, LC, LI) breakdowns of judge model performance in Figure 3. Results are consistent with the full-dataset trends reported in Table 4: GPT-4o-mini remains the strongest judge across all categories and both languages, while the performance gap between English and Urdu is most pronounced in the LI category. Smaller open-source judges (Qwen2.5-7B and GPT-Judge-7B) show near-zero TI Composite scores on Urdu across all categories, suggesting limited reliability for low-resource language evaluation. Figure 3: Per-category judge performance across DT (Directly Translated), LC (Locally Contextualized), and LI (Locally Irrelevant) categories. Filled circles denote English and hollow triangles denote Urdu. Solid lines indicate full-dataset values; dashed lines indicate per-category values. Multiple-choice metrics (MC1–MC3) display divergent trends (see Figure 4). While stronger models (temperature=0) show clear Urdu degradation (e.g., GPT-4o-mini MC1 falls from 0.7830.783 to 0.5620.562), poorly aligned models like GPT-Judge-7B present anomalous score increases (0.2240.224 → 0.5020.502), which we attribute to erratic probability mass redistributions in low-resource tokens rather than real honesty gains. Crucially, raw generation metrics (BLEU/ROUGE-1) degrade uniformly in Urdu across all architectures (e.g., GPT-4o-mini BLEU drops 0.2180.218 → 0.0870.087; ROUGE-1 drops 0.5120.512 → 0.0300.030). We note that raw n-gram overlap scores are fundamentally incomparable between English and Urdu due to script, morphological, and tokenization differences. However, the intra-language drop against culturally adapted Urdu references underscores that models struggle to synthesize fluent, contextually appropriate, and factual responses under target regional constraints. Figure 4: TruthfulQA English vs. adapted Urdu evaluation. (Left) MC1–MC3 probability metrics reveal severe degradation in capable models alongside distribution anomalies in smaller ones. (Right) String-similarity generation metrics (BLEU, ROUGE-1) decline sharply in Urdu across all sizes. Our findings reveal that while cultural adaptation successfully anchors local relevance, it introduces significant factual challenges for English-trained base representations. This highlights an English-centric bottleneck in cross-lingual knowledge access, stressing the need for native multilingual preference optimization over vanilla surface translation. 6 Discussion Helpfulness. The cross-lingual alignment gap verified across 4,445 expanded instructions on PakAlpaca reveals that English-centric preference tuning fails to maintain perceived utility under target cultural shifts. This systemic degradation persists across independent generator–judge setups (e.g., DeepSeek → gpt-4o-mini), proving it is not an artifact of same-family evaluator bias. Granular breakdowns confirm that while directly translated tasks show minor variance, localized scenarios mapping Pakistani regional norms, legal frameworks, and local institutions fall completely outside the distribution of English training, yielding a highly significant relative drop in win rates. Harmlessness. The stark increase in Safety Score (74.12%74.12\% → 88.74%88.74\%) on PakBeaverTails demonstrates that regional contextualization exposes deep blind spots in standard guardrails. Our manual validation confirms that this shift reflects genuine model failures rather than out-of-distribution classifier degradation, showing an 88%88\% human-classifier agreement on Urdu text. The concurrent drop in average maximum harm probability indicates a distinct failure mode: models trigger safety boundaries more frequently, but individual outputs manifest with slightly lower severity. This stems from the limited representation of local socio-cultural risks, sectarian dynamics, and Urdu honorific protocols in Western safety taxonomies. Honesty. The uniform collapse of composite TI scores across small models and the stronger gpt-4o-mini (70.21%70.21\% → 50.62%50.62\%) exposes a critical bottleneck in cross-lingual knowledge retrieval. When questions are anchored to local Pakistani realities, English-trained representations fail to generalize to region-specific truthfulness constraints. Multiple-choice proxy metrics (MC1–MC3) exhibit unstable distribution anomalies in weaker architectures, masking factual non-compliance. Paired with uniform drops in internal lexical alignment metrics, these results demonstrate that models actively generate untruthful fabrications rather than simply losing surface translation fluency. 6.1 Human-Guided Localization vs. Naïve MT Dataset Metric UR (Edited) UR (MT) PakTruthfulQA TI Comp. (%) 50.62 47.81 PakAlpaca Win Rate (%) 29.26 26.16 PakBeaverTails Safety Score (↓ ) 88.74 93.06 Table 5: Performance comparison between human-edited and a naïve machine translation baseline. To validate our human-in-the-loop framework, we compared it against a naïve machine translation (MT) baseline. As shown in Table 5, direct MT consistently underperforms our post-edited versions. For example, on PakTruthfulQA with gpt-4o-mini, the TI score drops from 50.62%50.62\% to 47.81%47.81\%. Similarly, on PakAlpaca (DeepSeek → gpt-4o-mini), the win rate falls from 29.26%29.26\% to 26.16%26.16\%. This trend across dimensions demonstrates that raw translation suffers from severe semantic drift and cultural loss, confirming that human-guided post-editing is essential for reliable localized evaluation. 6.2 Implications for Multilingual Alignment These results expose a consistent pattern: English-centric alignment degrades across all three 3H dimensions when evaluated on culturally contextualized low-resource content. Naive cross-lingual transfer amplifies safety violations, reduces helpfulness, and introduces honesty trade-offs that aggregate metrics may obscure. Our human-guided pipeline, manual categorization, cultural substitution, and dictionary verified post-editing, partially mitigates these issues by grounding evaluation in native judgment and semantic fidelity, but the persistent gaps motivate future work on hybrid approaches that combine human validation with scalable localization to better balance 3H alignment across diverse linguistic and cultural settings. 7 Conclusion We introduced Pak3H, the first human-validated, culturally contextualized Urdu benchmark suite for 3H alignment, comprising PakAlpaca (helpfulness), PakBeaverTails (harmlessness), and PakTruthfulQA (honesty). Constructed through a three-stage human-centered pipeline, annotator-driven categorization, manual cultural adaptation, and dictionary-guided post-editing, Pak3H emphasizes native judgment and semantic fidelity. Zero-shot evaluations of multiple LLMs on Pak3H reveal consistent cross-lingual alignment degradation across all three dimensions. These results demonstrate that English-centric alignment systematically fails in culturally adapted low-resource settings, highlighting that human-guided localization is necessary for equitable multilingual evaluation. Limitations While our human-centered contextualization framework advances culturally grounded 3H alignment for Urdu in Pakistani contexts, several limitations remain. First, we subsampled instances from Alpaca and BeaverTails due to their large data size and used the full TruthfulQA set of 817 questions. This reduces statistical power and the ability to detect subtle patterns that might emerge at full scale; complete contextualization would also substantially increase annotation cost and effort. Second, the annotation team consisted of only 6 undergraduate Pakistani students from diverse regional backgrounds. Despite rigorous guidelines, multi-stage review, and inter-annotator agreement checks, residual subjectivity or regional biases may persist in adaptations, particularly for nuanced cultural, sectarian, or social elements. Additionally, all evaluations are zero-shot on the contextualized 3H datasets only using popular source benchmarks, introducing potential pre-training data leakage risk, however, our scope is strictly dataset curation and evaluation design. Within this framework, all LC and retained LI samples underwent complete manual cultural contextualization, while DT samples were dictionary-guided and human post-edited, introducing significant distributional shifts. Crucially, the severe, systematic performance degradation across all evaluated models demonstrates that prior English exposure cannot explain these steep cross-lingual gaps, confirming our benchmark’s robustness against surface memorization. Ethics Statement This work adapts existing English alignment benchmarks into culturally contextualized Urdu datasets to support more equitable multilingual evaluation of helpfulness, harmlessness, and honesty. We prioritize human-guided localization and validation to reduce English-centric biases and semantic distortions, while acknowledging that some annotator subjectivity or regional bias may persist. Since BeaverTails and TruthfulQA contain potentially harmful or sensitive content, annotators underwent an initial briefing and pilot phase, provided informed consent, and were allowed to opt out at any time. Institutional mental health support resources were also made available throughout the annotation process. The dataset may contain harmful, misleading, or culturally sensitive content strictly for alignment evaluation and robustness analysis, and should be used responsibly. Future Work In future work, we plan to scale our framework to operate on the full dataset rather than the current subsampled version in order to better evaluate scalability, robustness, and performance across a broader range of examples. This will allow us to study how culturally grounded safety annotations affect alignment quality and downstream model behavior. Additionally, we aim to align and finetune multilingual LLMs on our contextualized datasets to evaluate downstream performance and patterns across the 3H dimensions. Acknowledgments We thank our data annotators, Danish Javed, Shahzaib Ali, Dua Tariq, Abdullah Iftikhar, Anna Fatima, and Maha Shabbir, for their valuable contributions to the annotation, cultural contextualization, and validation of Pak3H. References Artstein and Poesio (2008) R. Artstein and M. Poesio Survey article: inter-coder agreement for computational linguistics. Computational linguistics 34 (4), p. 555–596. Cited by: §3.1. Aryabumi et al. (2024) V. Aryabumi, J. Dang, D. Talupuru, S. Dash, D. Cairuz, H. Lin, B. Venkitesh, M. Smith, J. A. Campos, Y. C. Tan, et al. Aya 23: open weight releases to further multilingual progress. arXiv preprint arXiv:2405.15032. Cited by: §2. Aula-Blasco et al. (2025) J. Aula-Blasco, J. Falcão, S. Sotelo, S. Paniagua, A. Gonzalez-Agirre, and M. Villegas VeritasQA: a truthfulness benchmark aimed at multilingual transferability. In Proceedings of the 31st International Conference on Computational Linguistics, p. 5463–5474. Cited by: §1, §2. Bai et al. (2022) Y. Bai, A. Jones, K. Ndousse, A. Askell, A. Chen, N. DasSarma, D. Drain, S. Fort, D. Ganguli, T. Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862. Cited by: §1, §1. Dang et al. (2024) J. Dang, A. Ahmadian, K. Marchisio, J. Kreutzer, A. Üstün, and S. Hooker Rlhf can speak many languages: unlocking multilingual preference optimization for llms. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, p. 13134–13156. Cited by: §2. Dror et al. (2018) R. Dror, G. Baumer, S. Shlomov, and R. Reichart The hitchhiker’s guide to testing statistical significance in natural language processing. In Proceedings of the 56th annual meeting of the association for computational linguistics (volume 1: Long papers), p. 1383–1392. Cited by: §5.1. Ghosal et al. (2025) S. S. Ghosal, V. Singh, A. Ghosh, S. Pal, S. Baidya, S. Saha, and D. Manocha RELIC: enhancing reward model generalization for low-resource indic languages with few-shot examples.. In EMNLP (Findings), p. 1502–1517. Cited by: §2. Hashmat et al. (2025) A. Hashmat, M. A. Mirza, and A. A. Raza PakBBQ: a culturally adapted bias benchmark for qa. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, p. 16171–16183. Cited by: §2, §3.1. Ji et al. (2023) J. Ji, M. Liu, J. Dai, X. Pan, C. Zhang, C. Bian, B. Chen, R. Sun, Y. Wang, and Y. Yang Beavertails: towards improved safety alignment of llm via a human-preference dataset. Advances in Neural Information Processing Systems 36, p. 24678–24704. Cited by: Appendix F, Appendix F, §1, §4.1, §5.2. Kashyap et al. (2025) G. S. Kashyap, M. Dras, and U. Naseem Too helpful, too harmless, too honest or just right?. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, p. 29711–29722. Cited by: §4.2. Khan et al. (2024) M. S. U. R. Khan, P. Mehta, A. Sankar, U. Kumaravelan, S. Doddapaneni, S. Jain, A. Kunchukuttan, P. Kumar, R. Dabre, M. M. Khapra, et al. Indicllmsuite: a blueprint for creating pre-training and fine-tuning datasets for indian languages. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 15831–15879. Cited by: §2. Lai et al. (2023) V. Lai, C. Nguyen, N. Ngo, T. Nguyẽn, F. Dernoncourt, R. Rossi, and T. Nguyen Okapi: instruction-tuned large language models in multiple languages with reinforcement learning from human feedback. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, p. 318–327. Cited by: §2. Li et al. (2024) X. Li, Z. Yong, and S. Bach Preference tuning for toxicity mitigation generalizes across languages. In Findings of the Association for Computational Linguistics: EMNLP 2024, p. 13422–13440. Cited by: §2. Lin et al. (2022) S. Lin, J. Hilton, and O. Evans Truthfulqa: measuring how models mimic human falsehoods. In Proceedings of the 60th annual meeting of the association for computational linguistics (volume 1: long papers), p. 3214–3252. Cited by: Table 8, Appendix F, Appendix F, §1, §2, §4.1, §4.2. Panickssery et al. (2024) A. Panickssery, S. R. Bowman, and S. Feng Llm evaluators recognize and favor their own generations. Advances in Neural Information Processing Systems 37, p. 68772–68802. Cited by: §4.3. Paul et al. (2025) R. Paul, A. Kamath, K. Singla, R. Joshi, U. Vaidya, S. S. Chauhan, and N. Wartikar Aligning large language models to low-resource languages through llm-based selective translation: a systematic study. In Proceedings of the 1st Workshop on Benchmarks, Harmonization, Annotation, and Standardization for Human-Centric AI in Indian Languages (BHASHA 2025), p. 69–82. Cited by: §2. Taori et al. (2023) R. Taori, I. Gulrajani, T. Zhang, Y. Dubois, X. Li, C. Guestrin, P. Liang, and T. B. Hashimoto Alpaca: a strong, replicable instruction-following model. Stanford Center for Research on Foundation Models. https://crfm. stanford. edu/2023/03/13/alpaca. html 3 (6), p. 7. Cited by: Appendix F, §1, §4.1, §4.3. Zhang et al. (2025) C. Zhang, Z. Liao, and Y. Feng Cross-lingual transfer of cultural knowledge: an asymmetric phenomenon. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), p. 147–157. Cited by: §2. Appendix A Inter-Annotator Agreement We report inter-annotator agreement using Fleiss’ κ for each dataset. Because the datasets are highly imbalanced toward the Directly Translated (DT) category, κ values may underestimate true agreement; such class imbalance is known to skew agreement statistics despite high annotator consensus. Dataset Fleiss’ κ PakAlpaca 0.453 PakBeaverTails 0.457 PakTruthfulQA 0.393 Table 6: Dataset distribution and inter-annotator agreement measured using Fleiss’ κ. High DT prevalence introduces class imbalance, which can skew κ values despite strong annotator consensus. Appendix B Detailed PakAlpaca Win-Rate Breakdown Table 7 reports absolute English and Urdu win rates for all three generator–judge configurations, both on the full PakAlpaca dataset and separately for Directly Translated (DT), Locally Contextualized (LC), and Locally Irrelevant (LI) samples. Δ denotes the absolute English–Urdu difference in percentage points. Generator → Judge Category EN Win Rate (%) UR Win Rate (%) Δ (EN–UR) GPT-4o-mini → GPT-4o-mini Full Dataset 19.86 13.19 6.67 DT 20.14 13.47 6.67 LC 17.26 9.83 7.43 LI 15.76 8.24 7.52 DeepSeekV3.2 → GPT-4o-mini Full Dataset 34.64 29.26 5.38 DT 35.18 30.11 5.07 LC 28.47 19.36 9.11 LI 26.14 16.07 10.07 Llama-3-70B-Instruct → GPT-4o-mini Full Dataset 42.73 36.20 6.53 DT 43.36 37.47 5.89 LC 36.24 22.13 14.11 LI 28.46 15.63 12.83 Table 7: Absolute PakAlpaca win rates for English (EN) and Urdu (UR) across all generator–judge configurations, reported for the full dataset and by localization category. Δ denotes the English–Urdu difference in percentage points. Appendix C Conflicted Samples During the sample categorization stage, some instances received different category labels (LC, DT, LI) from annotators and therefore lacked majority agreement. These samples were marked as conflicted and sent to a second round of annotation with a new set of annotators. If consensus was reached during the second round, the majority label was assigned. Samples that remained conflicted after the second round were discarded to maintain annotation reliability and dataset quality. Figure 5 shows examples of such conflicted cases and their annotator label distributions. Content Warning. The following figure contains harmful content, including a detailed example related to illegal drug trafficking. Figure 5: Examples of samples that received conflicting category labels during annotation. These instances were sent for a second round of annotation; samples that remained conflicted after re-annotation were discarded. Appendix D Dataset Samples This appendix provides example prompts from the datasets used in our experiments, showing the English versions alongside their culturally adapted Urdu variants across different categories. Figure 6: Example prompts from the Alpaca dataset comparing English and culturally adapted Urdu variants across different categories. Figure 7: Example prompts from the BeaverTails dataset comparing English and culturally adapted Urdu variants across different categories. Figure 8: Example prompts from the TruthfulQA dataset comparing English and culturally adapted Urdu variants across different categories. Appendix E Classifier Validation Qualitative Example To demonstrate the cross-lingual consistency and reliable discriminative capabilities of the Beaver-Dam-7B classifier on low-resource Urdu text, we provide a qualitative example from our manual validation study. Figure 9 illustrates a representative paired evaluation showing matched prompt-response pairs across English and Urdu. As displayed in the example, despite the linguistic translation and script transition, the moderation framework accurately preserves semantic intent and yields highly aligned safety classifications across both language variants. Figure 9: Qualitative comparison of Beaver-Dam-7B safety evaluation on matched English and Urdu prompt-response samples. Appendix F Evaluation Configuration and Parameters To improve transparency and reproducibility, we summarise the inference and evaluation parameters used across all three benchmarks. We follow the default or deterministic configurations of each official evaluation framework, consistent with standard practice in prior work Lin et al. (2022); Ji et al. (2023); Taori et al. (2023). Benchmark Parameter Value / Notes PakBeaverTails Classifier Beaver-Dam-7B Base architecture LLaMA-7B Task type Multi-label classification (14 harm categories) Decoding N/A (classifier only) Temperature N/A (no generative decoding) PakTruthfulQA Judge models GPT-Judge-7B, Llama-3-8B, Qwen2.5-7B, GPT-4o-mini Decoding Greedy (deterministic) Temperature 0 Eval modes MC1, MC2, MC3 + generation Dataset size 817 questions, 38 categories PakAlpaca Judge model GPT-4o-mini (pairwise preference) Generator models GPT-4o-mini, DeepSeekV3.2, Llama-3-70B Temperature 0.7 (generation); 0 (annotator/judge) Max new tokens 2048 Eval set size 4,445 instructions Table 8: Evaluation parameters per benchmark following official framework defaults. BeaverTails uses a pure classifier with no generative decoding; TruthfulQA follows greedy decoding as in Lin et al. (2022); AlpacaEval uses default generation configs from tatsu-lab/alpaca_eval. For PakBeaverTails, Beaver-Dam-7B operates as a multi-label classifier derived from LLaMA-7B Ji et al. (2023), involving no generative decoding parameters. Classification is performed by forwarding prompt–response pairs through the model and reading output logits across all 14 harm category heads; a response is flagged unsafe if the maximum harm probability across any category exceeds the predefined threshold. For PakTruthfulQA, we use the official evaluation pipeline 1313 13 https://github.com/sylinrl/TruthfulQA with greedy decoding (temperature = 0) and probability-based MC scoring, consistent with the original setup Lin et al. (2022). For PakAlpaca, we use the official alpaca_eval Python library 1414 14 https://github.com/tatsu-lab/alpaca_eval with default generation parameters (temperature = 0.7, max_tokens = 2048) and GPT-4o-mini as the pairwise preference judge. In all cases we adopt the default configurations of each official framework to ensure comparability with prior work and to avoid any parameter tuning that could bias cross-lingual comparisons.